Big Data in Education

Alex J. Bowers, Ph.D.

Associate Professor of Education Leadership

Teachers College, Columbia University

April 24, 2014

Schutt, R., & O'Neil, C. (2013). Doing Data Science:

Straight Talk from the Frontline. Cambridge, MA:

O'Reilly.

Schutt, R., & O'Neil, C. (2013). Doing Data Science:

Straight Talk from the Frontline. Cambridge, MA:

O'Reilly.

You’re dealing with Big Data when you’re

working with data that doesn’t fit into your

computer unit. Note that makes it an evolving

definition: Big Data has been around for a

long time. . . . Today, Big Data means

working with data that doesn’t fit in one

computer. p.24

“Big” Analysis in Education Leadership

Schutt, R., & O'Neil, C. (2013). Doing Data Science:

Straight Talk from the Frontline. Cambridge, MA:

O'Reilly.

You’re dealing with Big Data when you’re

working with data that doesn’t fit into your

computer unit. Note that makes it an evolving

definition: Big Data has been around for a

long time. . . . Today, Big Data means

working with data that doesn’t fit in one

computer. p.24

A data scientist is someone who knows how

to extract meaning from and interpret data,

which requires both tools and methods from

statistics and machine learning, as well as

being human… Once she gets the data into

shape, a crucial part is exploratory data

analysis, which combines visualization and

data sense… She’ll communicate with

team members, engineers, and leadership in

clear language and with data visualizations

so that even if her colleagues are not

immersed in the data themselves, they will

understand the implications. p.16

Statistician

Data Scientist

http://flowingdata.com/2013/12/18/data-scientist-surpasses-statistician-on-google-trends/

December 18, 2013

Data scientist surpasses statistician on Google

Trends

Mike Bostock & D3:

github.com/mbostock

github.com/mbostock/d3/wiki/Gallery

Kris Erickson http://ericksondata.com/wp/2012/venn-diagram-of-data-science/

But what do school practitioners (the users)

actually say that they need?

But what do school practitioners (the users)

actually say that they need?

Brocato, K., Willis, C., Dechert, K. (2014) Longitudinal Data Use: Ideas for

District, Building and Classroom Leaders. In A.J. Bowers, A.R. Shoho, B. G.

Barnett (Eds.) Using Data in Schools to Inform Leadership and Decision

Making (forthcoming). Charlotte, NC: Information Age Publishing Inc.

Brocato, K., Willis, C., Dechert, K. (2014) Longitudinal

Data Use: Ideas for District, Building and Classroom

Leaders. In A.J. Bowers, A.R. Shoho, B. G. Barnett

(Eds.) Using Data in Schools to Inform Leadership and

Decision Making (forthcoming). Charlotte, NC:

Information Age Publishing Inc.

Brocato, K., Willis, C., Dechert, K. (2014) Longitudinal

Data Use: Ideas for District, Building and Classroom

Leaders. In A.J. Bowers, A.R. Shoho, B. G. Barnett

(Eds.) Using Data in Schools to Inform Leadership and

Decision Making (forthcoming). Charlotte, NC:

Information Age Publishing Inc.

What components of a statewide longitudinal

data system are needed for that system to best

meet the respective needs of superintendent,

principal and teacher leaders?

Superintendents

Teachers Principals

Longitudinal Data System

What components of a longitudinal data system are needed for that system

to best meet the respective needs of superintendent, principal, and teacher

leaders?

Superintendents (n=55)

Individual Student/Teacher

• Teacher data: performance, attendance, past experience

• Who rides the bus

• Record of parent conferences

• Test scores for students as well as whether or not they have met growth

Comparative

• All district’s comparative test data, district demographics

• Access test data, budget information, comparing per pupil expenditures, demographic data for the local area as compared to the region and state

• Student growth from year to year, demographic group growth

Principals (n=45)

Teacher Data

• Teachers’ past performance and years of experience

• Background of teachers and records showing student test scores broken down by teacher

• Student and teacher achievement and growth data, teacher evaluation information, student attendance and dropout data, staff attendance, accountability information

Student Data

• Achievement information on teachers based on student performance

• Background of teachers being hired; teacher performance data

• Are we able to keep high performing teachers in our district?

Teachers (n=193)

Brocato, Willis & Dechert (2014) p.5

Brocato, Willis & Dechert (2014) p.6

Teachers continued

Figure 1. Summary of eight themes and

corresponding frequency percentage in the

“more frequent” level of category

What components of a longitudinal data system are needed for that system

to best meet the respective needs of superintendent, principal, and teacher

leaders?

Which stages of the Data Science Process do we do well in education?

Where are the challenges?

Schutt, R., & O'Neil, C. (2013). Doing Data Science: Straight Talk from the Frontline. Cambridge, MA: O'Reilly.

K-12 Schooling Outcomes

• Dropout risk

– Accuracy, precision, sensitivity & specificity

• Longitudinal patterns in teacher assigned grades

• Data driven decision making (3DM)

• Targeted and data-informed resource allocation

to improve schooling outcomes

Bowers, A.J., Sprott, R., Taff, S.A. (2013) Do we Know Who Will Drop Out? A Review of

the Predictors of Dropping out of High School: Precision, Sensitivity and Specificity. The

High School Journal.96(2), 77-100. http://hdl.handle.net/10022/AC:P:21258

Issues of Accuracy in Predicting High School Dropout

• Dropping out of high school in the U.S. is associated with a multitude of negative outcomes.

• However, other than a select number of demographic and background variables, we know little about the accuracy of current dropout predictors that are school-related.

• Current predictions accurately predict only about 50-60% of students who actually dropout.

• According to Gleason & Dynarski (2002), accurate prediction of who will dropout is a resource and efficiency issue:

– A large percentage of students are mis-identified as at-risk.

– A large percentage of students who are at-risk are never identified.

• Current reporting of dropout “flags” across the literature is haphazard. – Almost none report accuracy

– Many report specificity or sensitivity, but rarely both

• A dropout flag may be highly precise, in that almost all of the students with the flag dropout, but may not be accurate since the flag may identify only a small proportion of the dropouts.

Balfanz et al. (2007); Bowers (2010); Gleason & Dynarski (2002); Pallas (2003); Rumberger (2004)

Re-analyzing Past Dropout Flags for Accuracy,

Precision, Sensitivity and Specificity

• Literature search. • Queried multiple databases:

– JSTOR, Google Scholar, EBSCO, Educational Full Text Wilson Web

• Studies were included that: – Were published since 1979 – Examined High School dropout – Examined school-wide characteristics and included all students – A focus on the student level – Reported frequencies for re-analysis

• Initially yield 6,434 overlapping studies – 301 studies were read in full – 140 provided school-wide samples and quantifiable data – 36 articles provided enough data for accuracy re-calculations – Yield 110 separate dropout flags

• Relative Operating Characteristic (ROC) – Hits versus False-Alarms

Dropout Graduate

Dropout a

True-positive

(TP)

Correct

b

False-positive

(FP)

Type I Error

a+b

Graduate c

False-negative

(FN)

Type II Error

d

True-negative

(TN)

Correct

c+d

a+c b+d a+b+c+d=N

Event

Pre

dic

tor

Precision = a/(a + b) Positive Predictive Value

True-Positive Proportion = a/(a + c) Sensitivity

True-Negative Proportion = d/(b + d) Specificity

False-Positive Proportion = b/(b + d) 1-Specificity

Event table for calculating dropout contingency proportions

Fawcett (2004); Hanley & McNeil (1982); Swets (1988); Zwieg & Campbell (1993)

“Hits”

“False Alarms”

0

0.2

0.4

0.6

0.8

1.0

Tru

e-p

ositiv

e p

roport

ion (

Sensitiv

ity)

0 0.2 0.4 0.6 0.8 1.0

False-positive proportion (1-Specificity)

Perfect prediction

Better prediction

Worse prediction

Low attendance

An example of the true-positive proportion plotted against the false-positive

proportion for Balfanz et al. (2007) comparing the relative operating characteristics

(ROC) of each dropout flag.

Bowers, A.J., Sprott, R., Taff, S.A. (2013)

0

0.2

0.4

0.6

0.8

1.0

Tru

e-p

ositiv

e p

roport

ion (

Sensitiv

ity)

0 0.2 0.4 0.6 0.8 1.0

False-positive proportion (1-Specificity)

Relative operating characteristics (ROC) of all dropout flags reviewed, plotted

as the true-positive proportion against the false-positive proportion. Numbers

refer to dropout indicator IDs.

Bowers, A.J., Sprott, R., Taff, S.A. (2013)

0

0.2

0.4

0.6

0.8

1.0

Tru

e-p

ositiv

e p

roport

ion (

Sensitiv

ity)

0 0.2 0.4 0.6 0.8 1.0

False-positive proportion (1-Specificity)

Relative operating characteristics (ROC) of all dropout flags reviewed, plotted

as the true-positive proportion against the false-positive proportion. Numbers

refer to dropout indicator IDs.

Bowers, A.J., Sprott, R., Taff, S.A. (2013)

0

0.2

0.4

0.6

0.8

1.0

Tru

e-p

ositiv

e p

roport

ion (

Sensitiv

ity)

0 0.2 0.4 0.6 0.8 1.0

False-positive proportion (1-Specificity)

Relative operating characteristics (ROC) of all dropout flags reviewed, plotted

as the true-positive proportion against the false-positive proportion. Numbers

refer to dropout indicator IDs.

Bowers, A.J., Sprott, R., Taff, S.A. (2013)

http://itp.wceruw.org/documents/KnowlesDropoutEarlyWarningSystemforITPJan2013.pdf

Knowles, J. (2014)

Knowles, J. (2014)

http://itp.wceruw.org/documents/KnowlesDropoutEarlyWarningSystemforITPJan2013.pdf